将Tensorflow对象检测与keras CNN分类器集成

Question

My problem is to detect object using mobilenet SSD and then to read data's from the bounding boxes using CNN classifier trained in Keras.everytime after the bounding box is obtained it has to be evaluated with CNN classifier. On referring number of questions in git and stack i started write my own coding.however after resetting tf graph using tf.reset_default_graph() and on loading .h5 from keras it throws an error as

"ValueError: 
 Tensor("loss/dense_6_loss/Const:0", shape=(), 
 dtype=float32) must be from the same graph as 
 Tensor("truediv_19:0", shape=(?, 36), 
 dtype=float32)."

I am doing instance detect and image retrieval task by Keras and Tensorflow as backend.

show:ValueError: tensor a must be from the same graph as tensor b.

The code is as below:

Merge.py
from keras import backend as K

g1=tf.Graph()
g2=Graph()
sess1=tf.Session(graph=g1)
sess2=Session(graph=g2)

def intiMaskrcnn():
     with g1.as_default():
          with sess1.as_default():
               Model1=........
tf.rest_defaut_graph()
def instanceDetect():
     K.set_session(sess1)
     with g1.as_default():
           Model1.predit()
            ............
k.clear_session()

def intiMobilenet():
    with g2.as_default():
         with sess2.as_default():
              Model2=........

def Retrieval():
    K.set_session(sess2)
     with g2.as_default():
         Model2.predit()
           ............

i need to know is it possible to integrate the tf and Keras in one pipeline simultaneously..if possible how?Thaks in advance

Answer 1

This helps me to run the keras model inside the tensorflow SSD object Detection.

* Use Seperate Graph.
* Use Seperate Session.
* reset the default Graph using tf.reset_default_graph()
* for keras Use K.clear

from keras.models import load_model
with tf.Session(graph=K.get_session().graph) as session:
    session.run(tf.global_variables_initializer())
    model = load_model('model.h5')
    predictions = model.predict(input)